Sakana AI launches Fugu Ultra v2 agent orchestrator

Sakana AI has released Fugu Ultra v2 on OpenRouter, introducing a learned orchestration system designed to coordinate multi-step reasoning, autonomous research, and software development. Instead of relying on a single monolithic model, the system uses a controller language model to decompose user requests into subtasks and route them to a pool of specialist agents. This pool can be dynamically updated as model pricing and availability shift. Notably, Sakana achieved its performance without including Fable 5, Fable 5.1, or GPT-6-Astra in its agent pool.

According to Sakana, Fugu Ultra v2 achieved the best or joint-best results on five of eight benchmarks. It scored 74.3 on DeepSWE, which measures real-world software engineering capabilities, and 48.3 on Chartography, a visual and structured-data reasoning test. This Chartography score significantly outpaced Opus 5, which scored 27.3, and Fable 5, which scored 29.5. The system also led on the SWEFish, GDP.pdf, and Toolathon benchmarks.

Available exclusively through OpenRouter, Fugu Ultra v2 is priced at $5 per million input tokens and $30 per million output tokens, though prompts exceeding 272,000 tokens incur a higher rate. The system features a 1-million-token context window and an August 2026 knowledge cutoff, supplemented by built-in web search. It offers an OpenAI-compatible API, supports tool calling, structured outputs, and PDF or image inputs. Developers can configure the system's reasoning effort using high, xhigh, or max settings to balance quality against latency and token consumption.

For practitioners, this architecture shifts the integration focus from individual model APIs to a managed routing endpoint. While the design reduces dependency on specific model providers, developers must evaluate the total task cost and latency, as internal orchestration tokens are billed at standard rates. Sakana also offers Fugu Max, a sibling configuration optimized for cost-to-quality efficiency, whereas Ultra v2 prioritizes maximum capability on complex tasks.